Skip to content


Repository files navigation


GHA Linux GHA OSX GHA Windows Documentation Doxygen -> gh-pages Binder Join the Gitter Chat

Multi-dimensional arrays with broadcasting and lazy computing.


xtensor is a C++ library meant for numerical analysis with multi-dimensional array expressions.

xtensor provides

  • an extensible expression system enabling lazy broadcasting.
  • an API following the idioms of the C++ standard library.
  • tools to manipulate array expressions and build upon xtensor.

Containers of xtensor are inspired by NumPy, the Python array programming library. Adaptors for existing data structures to be plugged into our expression system can easily be written.

In fact, xtensor can be used to process NumPy data structures inplace using Python's buffer protocol. Similarly, we can operate on Julia and R arrays. For more details on the NumPy, Julia and R bindings, check out the xtensor-python, xtensor-julia and xtensor-r projects respectively.

xtensor requires a modern C++ compiler supporting C++14. The following C++ compilers are supported:

  • On Windows platforms, Visual C++ 2015 Update 2, or more recent
  • On Unix platforms, gcc 4.9 or a recent version of Clang


Package managers

We provide a package for the mamba (or conda) package manager:

mamba install -c conda-forge xtensor

Install from sources

xtensor is a header-only library.

You can directly install it from the sources:

cmake -DCMAKE_INSTALL_PREFIX=your_install_prefix
make install

Installing xtensor using vcpkg

You can download and install xtensor using the vcpkg dependency manager:

git clone
cd vcpkg
./vcpkg integrate install
./vcpkg install xtensor

The xtensor port in vcpkg is kept up to date by Microsoft team members and community contributors. If the version is out of date, please create an issue or pull request on the vcpkg repository.

Trying it online

You can play with xtensor interactively in a Jupyter notebook right now! Just click on the binder link below:


The C++ support in Jupyter is powered by the xeus-cling C++ kernel. Together with xeus-cling, xtensor enables a similar workflow to that of NumPy with the IPython Jupyter kernel.



For more information on using xtensor, check out the reference documentation


xtensor depends on the xtl library and has an optional dependency on the xsimd library:

xtensor xtl xsimd (optional)
master ^0.7.5 ^11.0.0
0.25.0 ^0.7.5 ^11.0.0
0.24.7 ^0.7.0 ^10.0.0
0.24.6 ^0.7.0 ^10.0.0
0.24.5 ^0.7.0 ^10.0.0
0.24.4 ^0.7.0 ^10.0.0
0.24.3 ^0.7.0 ^8.0.3
0.24.2 ^0.7.0 ^8.0.3
0.24.1 ^0.7.0 ^8.0.3
0.24.0 ^0.7.0 ^8.0.3
0.23.x ^0.7.0 ^7.4.8
0.22.0 ^0.6.23 ^7.4.8

The dependency on xsimd is required if you want to enable SIMD acceleration in xtensor. This can be done by defining the macro XTENSOR_USE_XSIMD before including any header of xtensor.


Basic usage

Initialize a 2-D array and compute the sum of one of its rows and a 1-D array.

#include <iostream>
#include "xtensor/xarray.hpp"
#include "xtensor/xio.hpp"
#include "xtensor/xview.hpp"

xt::xarray<double> arr1
  {{1.0, 2.0, 3.0},
   {2.0, 5.0, 7.0},
   {2.0, 5.0, 7.0}};

xt::xarray<double> arr2
  {5.0, 6.0, 7.0};

xt::xarray<double> res = xt::view(arr1, 1) + arr2;

std::cout << res;


{7, 11, 14}

Initialize a 1-D array and reshape it inplace.

#include <iostream>
#include "xtensor/xarray.hpp"
#include "xtensor/xio.hpp"

xt::xarray<int> arr
  {1, 2, 3, 4, 5, 6, 7, 8, 9};

arr.reshape({3, 3});

std::cout << arr;


{{1, 2, 3},
 {4, 5, 6},
 {7, 8, 9}}

Index Access

#include <iostream>
#include "xtensor/xarray.hpp"
#include "xtensor/xio.hpp"

xt::xarray<double> arr1
  {{1.0, 2.0, 3.0},
   {2.0, 5.0, 7.0},
   {2.0, 5.0, 7.0}};

std::cout << arr1(0, 0) << std::endl;

xt::xarray<int> arr2
  {1, 2, 3, 4, 5, 6, 7, 8, 9};

std::cout << arr2(0);



The NumPy to xtensor cheat sheet

If you are familiar with NumPy APIs, and you are interested in xtensor, you can check out the NumPy to xtensor cheat sheet provided in the documentation.

Lazy broadcasting with xtensor

Xtensor can operate on arrays of different shapes of dimensions in an element-wise fashion. Broadcasting rules of xtensor are similar to those of NumPy and libdynd.

Broadcasting rules

In an operation involving two arrays of different dimensions, the array with the lesser dimensions is broadcast across the leading dimensions of the other.

For example, if A has shape (2, 3), and B has shape (4, 2, 3), the result of a broadcasted operation with A and B has shape (4, 2, 3).

   (2, 3) # A
(4, 2, 3) # B
(4, 2, 3) # Result

The same rule holds for scalars, which are handled as 0-D expressions. If A is a scalar, the equation becomes:

       () # A
(4, 2, 3) # B
(4, 2, 3) # Result

If matched up dimensions of two input arrays are different, and one of them has size 1, it is broadcast to match the size of the other. Let's say B has the shape (4, 2, 1) in the previous example, so the broadcasting happens as follows:

   (2, 3) # A
(4, 2, 1) # B
(4, 2, 3) # Result

Universal functions, laziness and vectorization

With xtensor, if x, y and z are arrays of broadcastable shapes, the return type of an expression such as x + y * sin(z) is not an array. It is an xexpression object offering the same interface as an N-dimensional array, which does not hold the result. Values are only computed upon access or when the expression is assigned to an xarray object. This allows to operate symbolically on very large arrays and only compute the result for the indices of interest.

We provide utilities to vectorize any scalar function (taking multiple scalar arguments) into a function that will perform on xexpressions, applying the lazy broadcasting rules which we just described. These functions are called xfunctions. They are xtensor's counterpart to NumPy's universal functions.

In xtensor, arithmetic operations (+, -, *, /) and all special functions are xfunctions.

Iterating over xexpressions and broadcasting Iterators

All xexpressions offer two sets of functions to retrieve iterator pairs (and their const counterpart).

  • begin() and end() provide instances of xiterators which can be used to iterate over all the elements of the expression. The order in which elements are listed is row-major in that the index of last dimension is incremented first.
  • begin(shape) and end(shape) are similar but take a broadcasting shape as an argument. Elements are iterated upon in a row-major way, but certain dimensions are repeated to match the provided shape as per the rules described above. For an expression e, e.begin(e.shape()) and e.begin() are equivalent.

Runtime vs compile-time dimensionality

Two container classes implementing multi-dimensional arrays are provided: xarray and xtensor.

  • xarray can be reshaped dynamically to any number of dimensions. It is the container that is the most similar to NumPy arrays.
  • xtensor has a dimension set at compilation time, which enables many optimizations. For example, shapes and strides of xtensor instances are allocated on the stack instead of the heap.

xarray and xtensor container are both xexpressions and can be involved and mixed in universal functions, assigned to each other etc...

Besides, two access operators are provided:

  • The variadic template operator() which can take multiple integral arguments or none.
  • And the operator[] which takes a single multi-index argument, which can be of size determined at runtime. operator[] also supports access with braced initializers.


Xtensor operations make use of SIMD acceleration depending on what instruction sets are available on the platform at hand (SSE, AVX, AVX512, Neon).


The xsimd project underlies the detection of the available instruction sets, and provides generic high-level wrappers and memory allocators for client libraries such as xtensor.

Continuous benchmarking

Xtensor operations are continuously benchmarked, and are significantly improved at each new version. Current performances on statically dimensioned tensors match those of the Eigen library. Dynamically dimension tensors for which the shape is heap allocated come at a small additional cost.

Stack allocation for shapes and strides

More generally, the library implement a promote_shape mechanism at build time to determine the optimal sequence type to hold the shape of an expression. The shape type of a broadcasting expression whose members have a dimensionality determined at compile time will have a stack allocated sequence type. If at least one note of a broadcasting expression has a dynamic dimension (for example an xarray), it bubbles up to the entire broadcasting expression which will have a heap allocated shape. The same hold for views, broadcast expressions, etc...

Therefore, when building an application with xtensor, we recommend using statically-dimensioned containers whenever possible to improve the overall performance of the application.

Language bindings


The xtensor-python project provides the implementation of two xtensor containers, pyarray and pytensor which effectively wrap NumPy arrays, allowing inplace modification, including reshapes.

Utilities to automatically generate NumPy-style universal functions, exposed to Python from scalar functions are also provided.


The xtensor-julia project provides the implementation of two xtensor containers, jlarray and jltensor which effectively wrap julia arrays, allowing inplace modification, including reshapes.

Like in the Python case, utilities to generate NumPy-style universal functions are provided.


The xtensor-r project provides the implementation of two xtensor containers, rarray and rtensor which effectively wrap R arrays, allowing inplace modification, including reshapes.

Like for the Python and Julia bindings, utilities to generate NumPy-style universal functions are provided.

Library bindings


The xtensor-blas project provides bindings to BLAS libraries, enabling linear-algebra operations on xtensor expressions.


The xtensor-io project enables the loading of a variety of file formats into xtensor expressions, such as image files, sound files, HDF5 files, as well as NumPy npy and npz files.

Building and running the tests

Building the tests requires the GTest testing framework and cmake.

gtest and cmake are available as packages for most Linux distributions. Besides, they can also be installed with the conda package manager (even on windows):

conda install -c conda-forge gtest cmake

Once gtest and cmake are installed, you can build and run the tests:

mkdir build
cd build
cmake -DBUILD_TESTS=ON ../
make xtest

You can also use CMake to download the source of gtest, build it, and use the generated libraries:

mkdir build
cd build
make xtest

Building the HTML documentation

xtensor's documentation is built with three tools

While doxygen must be installed separately, you can install breathe by typing

pip install breathe sphinx_rtd_theme

Breathe can also be installed with conda

conda install -c conda-forge breathe

Finally, go to docs subdirectory and build the documentation with the following command:

make html


We use a shared copyright model that enables all contributors to maintain the copyright on their contributions.

This software is licensed under the BSD-3-Clause license. See the LICENSE file for details.